BEBR

FACULTY WORKING PAPER NO. 89-1616

Joint Tests of Non-Nested Models and General Error Specifications

The JAN 2 5 1990

lilinols of Uriana-Cha>r.paign

Anil K. Bern Michael McAleer M. Has hem Pesamn

College of Commerce and Business Administration Bureau of Economic and Business Research University of Illinois Urbana-Champaign

BEBR

FACULTY WORKING PAPER NO. 89-1616

College of Commerce and Business Administration

University of Illinois at Urb ana -Champaign

November 1989

Joint Tests of Non-Nested Models and General Error Specifications

Anil K. Bera Department of Economics University of Illinois

Michael McAleer

Department of Statistics

Australian National University

and

Institute of Social and Economic Research

Osaka University

M. Hashem Pesaran

Trinity College

Cambridge

and

Department of Economics

University of California, Los Angeles

The authors wish to thank Les Godfrey and Yuk Tse for helpful discussions, and Essie Maasoumi for valuable suggestions and comments. The first author is grateful for research support from the Research Board and the Bureau of Economic and Business Research of the University of Illinois, and the second author acknowledges the financial support of the Australian Research Grants Scheme and the Australian Research Council.

ABSTRACT This paper is concerned with joint tests of non-nested models and simultaneous departures from homoskedasticity, serial independence and normality of the disturbance terms. Locally equivalent alternative models are used to construct joint tests since they provide a convenient way to incorporate more than one type of departure from the classical conditions. The joint tests represent a simple asymptotic solution to the "pre-testing" problem in the context of non-nested linear regression models.

Key Words and Phrases: locally equivalent alternative models; non-normal errors; non-spherical errors; pre-testing problem.

Digitized by the Internet Archive

in 2011 with funding from

University of Illinois Urbana-Champaign

http://www.archive.org/details/jointtestsofnonn1616bera

1. Introduction In recent years a substantial literature has been developed for testing non- nested regression models. While the available procedures are now frequently used for both testing and modelling purposes, in many cases it would seem that the non-nested models are presumed to have disturbances satisfying the classical conditions of serial independence (I), homoskedasticity (H) and normality (N). In practice, while departures from the classical conditions occur quite frequently, it is not straightforward to modify the available test procedures to incorporate all the departures, especially non-normality (N) of the disturbances. Moreover, in the nested testing situation, most of the popular tests are "one-directional" in that they are designed to test against only a single alternative hypothesis, and in most cases the tests are valid only when the other standard assumptions are satisfied. Many researchers have found that the one-directional tests are not robust in the presence of other misspecifi cations (see Bera and Jarque (1982) and the references cited therein). In Section 2, the robustness of several well known tests for both nested and non- nested hypotheses is discussed briefly. In Section 3, we develop a procedure for testing non-nested models together with simultaneously checking the sphericality and normality of the disturbance terms. Locally equivalent alternative models are used to construct joint tests since they provide a convenient method for incorporating more than one type of departure from the classical conditions. The joint tests represent a simple asymptotic solution to the "pre-testing" problem in the context of non-nested linear regression models. Some concluding remarks are given in Section 4.

2. Robustness of Several Existing Tests In testing nested hypotheses, two kinds of situations can occur, namely undertesting and overtesting. When departures from the null hypothesis are multi-directional and a one-directional test is used, undertesting is said to

- 2 -

occur. On the other hand, in overtesting, the test statistic overstates the alternative hypothesis. In both undertesting and overtesting, a loss of power is to be expected. However, joint testing of several hypotheses is rarely performed in practice, so that inferences are generally affected by undertesting. By drawing upon some Monte Carlo results from Bera (1982) and Bera and Jarque (1982), we highlight the effects of undertesting and the subsequent non- robustness of a commonly used test of heteroskedasticity.

Three convenient simplifying assumptions are usually made in standard regression analysis, namely H, I and N. In what follows, we consider the Breusch and Pagan (1979) Lagrange multiplier (LM) test for heteroskedasticity (H). The data are generated under different combinations of N (the t distribution with five degrees of freedom), H (additive heteroskedasticity) and serial dependence (I) (first-order serial correlation); for further details, see Bera (1982) and Bera and Jarque (1982). The estimated powers of the test statistic under different data generating processes (DGP) are given below.

Nature of Alternative Model Correct Contaminated Misspecified

DGP HIN HIN HIN HIN HIN HIN HIN

Estimated power -726 -660 -270 -302 084 -222 128

When the DGP contains only heteroskedasticity (that is, HIN), the LM test is asymptotically optimal against the correct alternative and power is estimated to be -726. However, the estimated powers fall to -660 and -270 when the data are contaminated by the t distribution (HIN) and first-order serial correlation

(HIN), respectively. The effect of I on the LM test of H is quite substantial. When both I and N contaminate the DGP, the estimated power is only -302. The final three cases are completely misspecified in that the LM test is seeking to detect H when H does not exist but I, N or both I and N are present. The interpretation of the estimated powers depends on how the tests are to be

- o -

viewed. If the test is seen exclusively as a test of H, then the estimated powers should equal the estimated significance levels or sizes. It is clear from the table that the estimated sizes are significantly greater than the nominal size of the test in all three cases. However, if the test is used purely as a significance test, the estimated powers seem to be quite low, and are much lower than those for the case where the alternative model is contaminated. Regardless of the interpretation, therefore, the test does not perform well. .The one-directional test is not robust when the alternative model is contaminated and the test is not satisfactory in the completely misspecified case, regardless of how it is interpreted.

Turning now to the case of non-nested hypotheses, extensive Monte Carlo experiments have been conducted by Godfrey and Pesaran (1983) and Godfrey et al. (1988). The first of these two papers is concerned with the selection of regressors in two non-nested linear regression models, and examines the Cox test of Pesaran (1974), two mean- and variance-adjusted versions of the Cox test, the J test of Davidson and MacKinnon (1981), the JA test of Fisher and McAleer (1981), and the standard F test applied to the comprehensive model constructed as a union of the two models. Since the tests are valid only asymptotically when the disturbances are not normally distributed, Godfrey and Pesaran (1983) examine the robustness of the tests to errors drawn from the log-normal distribution and the chi-squared distribution with two degrees of freedom. Their experiments indicate that the finite sample significance levels are not significantly distorted and are broadly similar to those for the case of normally distributed errors. Although estimated powers tend to be greater when the errors are drawn from the two non-normal distributions compared with the normal case, the relative rankings of the tests in terms of power are not affected by the non-normality.

Various procedures are considered by Godfrey et al. (1988) for testing the non-nested linear and logarithmic functional forms. The test procedures are

- 4 -

classified as non-nested tests, two versions of the LM test and a variable addition test based on the more general Box-Cox transformation, and diagnostic tests of (possible) functional form misspecification against an unspecified alternative hypothesis. If the logarithmic model is to be taken seriously, the dependent variable of the linear model cannot take on negative values and the disturbances of the linear model cannot be normal. Therefore, it is essential to examine the robustness of the tests to non-normality of the errors, even if the primary consideration rests with testing the non-nested functional forms. Godfrey et al. (1988) examine the finite sample significance levels and powers of the tests when the disturbances follow the gamma (2,1) distribution, the log-normal distribution and the t distribution with five degrees of freedom. The two versions of the LM test based on the Box-Cox model are found to be highly sensitive to non-normality in that the estimated significance levels are far greater than those predicted by asymptotic theory, even when the sample size is eighty. On the other hand, the variable addition tests in the three categories are found to be robust to non-normality of the errors, and their relative rankings in terms of power are not affected by departures from normality.

3. Joint Tests

The standard situation for testing non-nested linear regression models with normal and spherical errors is as follows. It is desired to test the null model H0 against the non-nested alternative Hi, where the two models are

given as

2 Ho : y = X(3 + uo , uo ~ N(0,a0In)

and

2 Hi : y = Zy + ui , ux ~ N(0,a1In) ,

in which y is the n x 1 vector of observations on the dependent variable, X and Z are n x k and n x g matrices of observations on k and g linearly independent regressors, p and y are k x 1 and g x 1 vectors of unknown parameters, and uo and ui are vectors of normally, independently and identically distributed disturbances. It is also assumed that X and Z are not orthogonal, and that the limits of n^X'X, nrlZ'Z and n^X'Z exist, with the first two positive definite and the third non-zero. If X and Z contain stochastic rather than fixed elements, the probability limits of the appropriate matrices must exist, and X and Z must be distributed independently of uo and ui under Ho and Hi, respectively.

In considering the consequences of testing for certain departures from sphericality, it will be convenient to rewrite the two models as

Ho : yt = xt'p + uot (1)

and Hi : yt = zt'y + uu , (2)

in which xt' and zt are the t'th rows of X and Z, respectively, and t = l,2,...,n. When the assumptions regarding uot and uit are not satisfied, some of the properties of the tests will be affected. For example, Pesaran (1974) derived a test of non-nested linear regression models where the disturbances of each model follow a first-order autoregressive scheme. However, Pesaran's test is very complicated to apply in practice and a simpler procedure is given in McAleer, Pesaran and Bera (1989). The effect of heteroskedasticity will be similar. In particular, a straightforward application of the tests suggested in Davidson and MacKinnon (1981) and Fisher and McAleer (1981) will not be valid since the standard errors will not be correct. However, use of a heteroskedasticity-consistent covariance matrix estimator will circumvent this problem. When the errors are not normal, Pesaran's test based on the work of Cox (1961, 1962) is still valid asymptotically, although its small sample properties will be affected. Normality is required for the test suggested by

- 6 -

Fisher and McAleer (1981) to have the exact t-distribution under the null hypothesis; if normality does not hold, the test will be valid only asymptotically.

In the light of the above discussion, a basic requirement for applying standard non-nested testing procedures to achieve high power is that the models under consideration be well-specified. This means that tests for normality and sphericality, for example, are to be performed prior to testing the non-nested models themselves. An important, and frequently overlooked, aspect of testing non-nested models in this two-step procedure is the effect that such "pre-testing" may have on the levels of significance and powers of the non-nested tests. Therefore, it may be desirable to test the non-nested specifications jointly with departures from the classical assumptions regarding the disturbance terms. Such procedures will be particularly useful when there is a possibility of a non-normal disturbance term of unknown type, since it is frequently difficult to take account of general forms of non-normality in a straightforward manner. Joint tests may be constructed in a straightforward way by developing an approximate model which incorporates the various departures from the classical conditions into the systematic part of the model, so that the disturbances of the approximate model are normally, independently and identically distributed. The advantage of joint tests over the two-step testing procedure lies in the way the joint testing procedure deals with the "pre-testing" problem, at least asymptotically.

In the context of deriving Lagrange multiplier (LM) diagnostic tests,

Godfrey and Wickens (1982) suggested a way of obtaining a local approximation

to a given model with a non-standard disturbance structure. Such

approximations are called "locally equivalent alternative" (LEA) models. As

an illustration, consider Ho in (1), where it is now assumed that the

disturbance uot is given by

2 uot = Pouot-i + eot , eot ~ NED(0,a0) , I p0 1 < 1 (3)

- 7 -

for t = 2,3,.. .,n. A LEA model to Ho may be written as

3fc ^*

HQ : yt = xt'p + pouot-i + eot , (4)

in which uot = yt -xt'P and p is the ordinary least squares estimate of p under Ho.

The models HQ in (4) and Ho in (1) and (3) are "equivalent" in the sense that:

* (i) when po = 0, Ho and HQ are identical;

(ii) when p0 = 0, then 3it(P»^o»Po^Po = ^t^P,aO»Po^PO' wnere

* it(P>°2>Po) and ^f (P>a2>Po) are the log-density functions for

the t'th observation under Ho and HQ, respectively.

Godfrey (1981) has shown that, in testing po = 0 for local alternatives, the

likelihood ratio test of po = 0 applied to Ho in (1) and (3) and the LM test of po = 0

* applied to HQ in (4) have similar power. Thus, for values of po in the

* neighbourhood of zero, Ho and HQ may be regarded as equivalent.

Now let us consider (1) and (2) allowing for the possibility that the disturbances uit (i = 0,1) follow stationary autoregressive processes of order pj (i = 0,1), namely AR(pj), as follows:

Pi uit = I PijUit-j + Eit . i = 0,1 j=l

where t = p+l,p+2,...,n and p = max(po.pi). In this case, a locally equivalent form of (1) may be written as

* Po ~

HQ:yt = xt'p+ I pojuot-j + eot , j=i

where uot = yt - xt P» as before.

Although several procedures are available in the literature (for a recent

review, see McAleer and Pesaran (1986)), a convenient test of the null model

against both the non-nested alternative Hi and AR(po) disturbances can be performed by testing a = poi = P02 = ••• = Pop0 = 0 in the auxiliary linear

regression required for the J test of Davidson and MacKinnon (1981), namely

- 8 -

PO ^

yt = xt'p + I pojuot-j + ccyu + et , where yit is the predicted value of yt from (2), namely

•S •% Pi /N ^

yu = ztY+ I Pijuit-j .

uit = yt - ztY ana" Y is the maximum likelihood estimate of y under Hi.

An attractive feature of this approach is that other departures from the classical conditions may be handled in an equally straightforward manner. Consider the following general form of the distribution of the disturbance term for Ho of (1), where uot follows an autoregressive process of order po, namely

Po uot = X Pojuot-j + eot »

j=l

and eot is independently distributed. The density of Eot, denoted by g(eot)» is assumed to be a member of the symmetric Pearson family of distributions. This is not a very restrictive assumption since this family encompasses many distributions such as the normal, Student t and F. The density of e0't is then given by

g(eot) = exp pF(eot)l

J exp pF(eot)l deot

-oo < eot < oo

where

f 2

¥(eot) = J [ - eotAcot + cieot)] deot

When Ci = 0, g(eot) reduces to a normal density with mean zero and variance cot- Heteroskedasticity is introduced through cot- It is assumed that

cot = h

' qo n

i=l

where the elements of the qo x 1 vector vt = (vit,V2t,.--,vqot)' are fixed and measured around their means, h() is a twice differentiable function with h(0)

- y -

- o0, and (pi (i = l,2,...,qo) are unknown- parameters. Under these circumstances, the disturbances for Ho in (1) are now non-normal, heteroskedastic and serially dependent. A simple local approximation to this complicated model may be written as

in which

** Po ^ ~ Qo

HQ :yt = xt'(3+ X pojuot-j + uot I <Pi vit + Cirt + £ot , j=l i=l

rt = (uQt - 3uota0 )/ (4a0 ) ,

(5)

~2

n

~2

a0 = n-i I uQt,

L— A

with Eot ~ NID(0,a0) for all t.

To verify whether model (5) is a LEA model, we first note that, when po = (poi,p02,...,pOp0)' = 0, (p = ((pi,92,...,9q0)' = 0 and Ci = c2 = 0, H0 in (1) and HQ in (5) are identical and eot = uot> so that condition (i) is satisfied. Godfrey and Wickens (1982) verified condition (ii) above for the case of serial correlation and heteroskedasticity. It is, therefore, necessary to consider only the non-normal components. First, it can be shown that there is no contribution from the Jacobian term. The Jacobian from cirt is asymptotically given by (see Godfrey and Wickens (1982, p.85))

I in

t

3(u0t-o0)

1-ci

4o,

which, under local alternatives, reduces to

3C1 2 2.

--^Z(uot-G0),

4o0 t

which is 0p(l). Therefore, for the purpose of developing a joint test, we can ignore the Jacobian term.

It can also be shown that the score with respect to Ci is the same under Ho ** and HQ . From Bera and Jarque (1982, p.78), it follows that

3it(P.o0'0) uot

2 where ^tCP^o^) *s *he log-density function under Ho, with y =

2 (P0i,p02,-..,p0p0,9i,92,-.MCpq0,ci)'. Using the information that eot ~ NID(0,a0) for

all t, it follows from (5) that

dJtt (p,a0,0)

r = rt uot/a0 ,

dci

where Z ^ (•) is the log-density function under HQ . Since

n 9it(P,^,0) n 3it (P^o.0)

^ ic~i = ^ dci

t=l dCl t=l x

the score under the original and LEA models is the same. One component of

the LM test for normality is based on this score value (see Bera and Jarque

(1981), and Jarque and Bera (1987)).

If it were desired to test serial independence, homoskedasticity and normality under Ho, we would test the parameter restrictions

po = 0, 9 = 0, ci = 0 in equation (5). This joint test procedure has been suggested by Bera and Jarque (1982). However, they did not consider the possibility of the non-nested alternative Hi together with the non-sphericality and non-normality of the disturbance term under Ho-

If suitable predictions yit from a non-nested alternative Hi are augmented to equation (5) to yield the auxiliary regression given by

Po «- ~ qo ~

yt = xt'p+ £ pojuot-j + uot I 9i vit + cirt + ayn + eot , (6)

j=l i=l

then the null hypothesis, namely equation (1) with uot = £ot» involves a joint test of

H : po = 0, 9 = 0, ci = 0, a = 0. (7)

This joint hypothesis can be tested by applying the LM procedure, for example, directly to equation (6) or by using an appropriately adjusted F test (see Godfrey and Wickens (1982)). In computing the LM test, the most convenient form is nR2, that is, the sample size times the (uncentred) coefficient of determination in the auxiliary regression of a vector of ones on the following variables:

** dJt

~~* •%>■)

ap

= xt'uot/a0

**

^2 ~~2 ~4 2~= (UOt-<*oy(2CT0>

3g0

**

= uotuot-j/a0 , j = 1,2,. ..,po

3poj

**

aif

= (u^v^t/Oo) - vit , i = l,2,...,q0

a»i "VOt

** = (rtuot/o0) - 3(u0(.- a0)/(4a0)

and

**

"sr = yituot/°o •

The set of regressors given above can be simplified to

(xt'uot,uot- <yo»u°tuot-i,--,uotuot-p0,(uot- a0)vit,...,(uot- cJ0)vqot, 4rtuota0 - 3(uQt- a0), yituot).

As noted by Godfrey and Wickens (1982, p.86), it is not valid to apply the standard form of the F statistic to (6) in testing H in (7). For example, Godfrey and Wickens have shown that, for testing 9 = 0, the usual regression formula omits the factor 2 arising from the asymptotic distribution of

n^Zuot(uotvit)- Consider now the F statistic for testing ci = 0. The asymptotic t

variance of n-^ £ uotrt is the same as that of t

h-»I(u0t-3u0ta0)/(4a0).

2 Under H, uot = £ot ~ NID(0,a0) for all t. Therefore, the above variance can also

be expressed as

[E(u®t) - (E(uJt))2 + 9ao(E(uQt) - E(uot)2)

- 6G20(E(u0t) - E(uJt)E(uQt))]/(16ao)

= [(105oo - 9a®) + 9ao(3ao - aj) - 6o2Q(15ol - dafyo&tf

= 21aJ/8.

However, the limiting value of the regression formula is

a0 plim n-1 £ {\iQt- Bu^^dGa^ = 30^8 , t

which is one-seventh of the correct asymptotic variance. Denote the standard F statistics for testing

- 13 -

(po = 0, a = 0), 9 = 0 and ci = 0 by Fi, F2 and F3, respectively. Note that the regressors corresponding to the above parameters are asymptotically uncorrelated with each other, and also the regressors with coefficients 9 and ci are asymptotically orthogonal to the regressors of the null model. Therefore, the conditions for the decomposition of the joint test are satisfied (see Godfrey (1988, p.79)). It follows that, under Ho,

(po+DFi + K2q0F2 + V7F3 -d* X2(po+qo+2).

This test statistic will test the standard linear model (1) with normal and spherical disturbances against a broader alternative of non-spherical and non- normal disturbances, as well as against a non-nested alternative. Depending on the situation, it is possible to specialize the test statistic to particular alternatives by retaining the appropriate regressors in equation (6). For example, to test the null model Ho against a non-nested alternative Hi and the possible presence of heteroskedasticity, it would be necessary to test 9 = 0 and a = 0 in the auxiliary regression given by

yt = xt'p + uot I <Pivit + otyit + eot • i=l

This auxiliary regression equation is simply a specialization of equation (6)

with p0j = 0 (j=l,2,...,p0) and C\ = 0.

4. Conclusion In this paper we have presented some simple joint tests of non-nested models and general error specifications. The joint tests for non-nested specifications and for one or more departures from the classical conditions of serial independence, homoskedasticity and normality were developed within the context of locally equivalent alternatives. These tests represent a simple asymptotic solution to the "pre-testing" problem as applied to non-nested linear regression models. If the null hypothesis is not rejected by the joint tests,

- 14 -

standard regression analysis would follow for the underlying null model. However, if the null is rejected, it is not possible to infer whether it is rejected because of the non-nested alternative or through departures from the classical conditions regarding the disturbances.

References

Bera, A.K. (1982), Aspects of Econometric Modelling, Unpublished doctoral thesis, Australian National University.

Bera, A.K. and CM. Jarque (1981), An efficient large sample test for

normality of observations and regression residuals, Working Paper in Economics and Econometrics No. 040, Australian National University.

Bera, A.K. and CM. Jarque (1982); Model specification tests : A simultaneous approach, Journal of Econometrics, 20, 59-82.

Breusch, T.S. and A.R. Pagan (1979), A simple test for heteroscedasticity and random coefficient variation, Econometrica, 47, 1287-1294.

Cox, D.R. (1961), Tests of separate families of hypotheses, Proceedings of the Fourth Berkeley Symposium on Mathematical Statistics and Probability, 1 (Berkeley, University of California Press), 105-123.

Cox, D.R. (1962), Further results on tests of separate families of hypotheses, Journal of the Royal Statistical Society B, 24, 406-424.

Davidson, R. and J.G. MacKinnon (1981), Several tests for model specification in the presence of alternative hypotheses, Econometrica, 49, 781-793.

Fisher, G.R. and M. McAleer (1981), Alternative procedures and associated tests of significance for non-nested hypotheses, Journal of Econometrics, 16, 103-119.

Godfrey, L.G. (1981), On the invariance of the Lagrange multiplier test with respect to certain changes in the alternative hypothesis, Econometrica, 49, 1443-1455.

Godfrey, L.G. (1988), Misspecification Tests in Econometrics: The Lagrange Multiplier Principle and Other Approaches (Cambridge University Press, New York).

Godfrey, L.G., M. McAleer and CR. McKenzie (1988), Variable addition and Lagrange multiplier tests for linear and logarithmic regression models, Review of Economics and Statistics, 70, 492-503.

- 15 -

Godfrey, L.G. and M.H. Pesaran (1983), Tests of non-nested regression

models : Small sample adjustments and Monte Carlo evidence, Journal of Econometrics, 21, 133-154.

Godfrey, L.G. and M.R. Wickens (1982), Tests of misspecification using locally equivalent alternative models, in G.C. Chow and P. Corsi (eds), Evaluating the Reliability of Macroeconomic Models (Wiley, New York), 71-99.

Jarque, CM. and A.K. Bera (1987), An efficient large-sample test for normality of observations and regression residuals, International Statistical Review, 55, 163-172.

McAleer, M. and M.H. Pesaran (1986), Statistical inference in non-nested

econometric models, Applied Mathematics and Computation, 20, 271-311.

McAleer, M., M.H. Pesaran and A.K. Bera (1989), Alternative approaches to testing non-nested models with autocorrelated disturbances : An application to models of U.S. unemployment, Paper presented at the Time Series Conference, Osaka, Japan, June 1989.

Pesaran, M.H. (1974), On the general problem of model selection, Review of Economic Studies, 41, 153-171.